Tag Archives: linguistics

An unusual vowel contrast

Geoff Lindsey writes:

I do think that the official IPA chart is crowded with symbols which exist not so much for good acoustic or linguistic reasons as to fill the slots implied by its tongue-space framework.

For instance, I’m not sure that languages ever contrast ɨ and ɯ; the unrounded close non-front vowel of languages like Turkish and Vietnamese is sometimes transcribed as ɯ and sometimes as ɨ.

A while ago, I downloaded PHOIBLE‘s database, since they don’t have a search function on the site. Searching the database for inventories from the same source that contain both ɨ and ɯ, I get:

  • Apatani (RA)
  • Bora (SAPHON)
  • Kenyang (PH)
  • Kilba (GM)
  • Kod̩agu (RA)
  • Matses (PH)
  • Matsés (SAPHON)
  • Miraña (SAPHON)
  • Mishmi (RA)
  • Nimboran (UPSID)
  • Sedik (PH)
  • Sema (RA)
  • Southern Ute (PH)
  • Tangkul Naga (RA)
  • Wayana (SAPHON)

This is of course inconclusive, since the databases PHOIBLE draws from are unreliable. The Apatani vowel inventory sourced from RA in PHOIBLE differs wildly from the more usual one given in this paper.

However, this paywalled study claims that Bora really does have this contrast. SIL’s grammar of Kenyang claims that the language contrasts all of /ɨ ɯ u/, which is an even stronger example than Nimboran, the standard example of a language with such a contrast. (Nimboran has a six-vowel system with no rounded vowels, as does Matses.)

The Dravidian-Australian connection?

The indigenous languages of Australia have striking similarities in their phoneme inventories. Most have no fricatives, and none, as far as I know, have sibilants. Australian languages tend to have retroflexes (one exception is Bandjalang, which has only 12 consonants), palatals, contrastive dentals and alveolars, and several laterals and rhotics. Gasser and Bowern (2013) provide a chart of a ‘Standard Average Australian’ phoneme inventory:


Individual languages, of course, diverge from this pattern. Some have a voicing contrast in plosives, although this may really be a gemination contrast: Gasser and Bowern report that 59% of languages with a plosive voicing contrast collapse it initially. (Cf. the Proto-Basque fortis-lenis contrast: in plosives, this has become a voicing contrast, but fortis consonants could not appear word-initially and lenis consonants could not appear word-finally.) Over half have contrastive vowel length.

But there is much less difference across inventories in Australia than in most parts of the world. For example, Guugu Yimidhirr, the source of the English word ‘kangaroo’, differs only in having only one lateral (the apico-alveolar /l/) and a length contrast in its vowels. Dyirbal, the language whose noun class system inspired the title of George Lakoff’s Women, Fire, and Dangerous Things, is somewhat more divergent: its only retroflex consonant is the flap /ɽ/, its only lateral is the apico-alveolar /l/, and its only laminal consonant series is palatal.

What happens if we compare Standard Average Australian to Tamil?


The vowel inventory of Tamil differs significantly from that of SAA—Tamil has the mid vowels /e o/, the diphthongs /ai au/, and a length contrast in monophthongs—but the underlying consonant inventory is almost identical. The only significant differences are in the liquids: the rhotics are both retroflex, and Tamil has fewer laterals than SAA.

Looking to modern languages to try to gain information on ancient patterns is of course untenable as serious methodology, but it’s at least suggestive. Here’s Proto-Dravidian.


The velar nasal, present in all Australian languages in Gasser and Bowern’s sample, may not have been phonemic in Proto-Dravidian, but the general similarity is clear. I know of no other non-Australian language family that so closely follows the typical Australian language pattern, so, although I have no suggestion as to the causal relation here, no hypothetical scenario of ancient contact between India and Australia that could have spread this pattern from one place to the other, I suspect that one exists.

Interestingly, a ban on word-initial retroflexes is common among Australian languages, and Proto-Dravidian (like some languages of Western Australia, although I don’t know when this restriction developed) banned all apicals from word-initial position.

Intervocalic fortition

…was mentioned in xkcd:

But this has actually happened.

Blust’s compilation of highly unusual sound changes includes intervocalic fortition of *v *j *g in Kiput:



Intervocalic devoicing of *g also occurred in Berawan, which also reflects *b as k.

A possible explanation of both of these processes is here: Berawan -b- > β > ɣ, -g- > ɣ, -ɣ- > x > k; Kiput j- > d (leaving *j to only appear intervocalically), -g- > ɣ, and unconditional devoicing of the voiced fricatives and affricate. This doesn’t explain the intervocalic fortition of *w *y, however; they remained approximants at least word-finally.

A dialect of American English influenced by Pennsylvania Dutch has devoicing of plosives not only in coda position, but also in onset position in word-internal unstressed syllables.

Starostin claims “occasional intervocalic devoicing” for Dongxiang and Bonan.

Are there any instances of intervocalic fortition that aren’t devoicing? Kiput glide frication is probably an example, and Berawan (again) has conditioned gemination of intervocalic plosives:

Long Terawan examples such as *batu > bittoh “stone”, *kutu > kuttoh “head louse”, *qatay > atay “liver”, *putiq > puté “white”, *laki > lakkéh “man, male”, *siku > sikkoh “elbow”, *likud > likon “back (anat.)”, *tukud > tukon “prop, support”, *bana > binneh “husband”, *tina > tinneh “mother”, or *tanaq > tana “earth” show an unusual condition for the genesis of geminate consonants: the onset of an open final syllable was geminated. Although the data are more abbreviated, an identical change appears to be reflected in all Berawan dialects. Note that neither the syllable type nor its position are sufficient in themselves to predict gemination, as the consonant onsets of open penultimate syllables, or of closed final syllables remain unaffected. In citation forms stress is generally final in all dialects of Berawan, but this is true whether the final syllable is open or closed. What linguistic factor, if any, might drive consonant onsets to geminate only if they initiate an open final syllable thus remains very puzzling.

Doesn’t Italian have sporadic intervocalic gemination?


Reanalyzing the English vowel system

I thought I’d already written this post, but if I have, I can’t find it.

Traditionally, American English is analyzed as having eleven or twelve monophthongs, /æ ɑ ɛ ʌ ɔ e o ɪ ʊ i u/ + /ɚ/, and three diphthongs, /ai au oi/; but there’s a common dialectal feature that are difficult to explain in this analysis.

This process is l-breaking: after /e i ai oi/, /l/ breaks to /(j)əl/—so the words palepeelpile, and boil all have two syllables. L-breaking occurs sporadically after /au/—for me, howl has one syllable and owl has two—and some speakers have it after /u/, so school has two syllables for them.

In the traditional analysis, /e i ai oi/ don’t form a natural class. They’re all front vowels, and they’re all front ‘long vowels’, but what are long vowels really? Certainly not phonemically long; nobody believes that the contrast between /ɛ/ and /e/—or, worse, /ɛ/ and /i/—is one of length. It would be possible to posit a ‘long vowel’ category that only makes sense diachronically, if not for the fact that l-breaking also occurs after Vr sequences. 

The only Vr sequence (other than the rhotic vowel, which doesn’t trigger l-breaking) that occurs before /l/ in common English is /ar/, about which see here. (It’s sporadic in that show, but it can occur.) /erl/ can occur in some foreign surnames, and in my experience, it’s always pronounced with two syllables. I don’t know of any examples of /irl/ or /orl/.

How can /e i ai oi ar er/ (and maybe /ir or/, and sometimes /au/ and /u/) form a natural class? The answer is simple: phonetically, /e i ai oi au u/ are all diphthongs. (/o/ is also phonetically a diphthong [əu], but it has an allophone [o] that appears before /l/; [o] is in fact marginally phonemic for some speakers, since it occurs irregularly in a few words, like both, which can be seen misspelled bolth.) This suggests that Vr sequences should also be analyzed as diphthongs. The choice of symbol is somewhat arbitrary (although monophthongized /ai/ contrasts with both [æ] and [ɑ])—but in this analysis, American English has eight monophthongs, /æ ɑ ɛ ʌ ɔ ɪ ʊ ɚ/; four i-diphthongs, /ai ei oi ɪi/; three u-diphthongs, /æu əu ʊu/; and four ɚ-diphthongs, /aɚ eɚ oɚ iɚ/.

This analysis also explains some of the phonetic drift in some dialects: the first components of /əu ʊu/ front because there are no diphthongs like /eu iu/ for them to contrast with, and the first components of /oi oɚ/ raise for similar reasons.

A similar analysis of British English is here.

Person-marking in Tangut

Guillaume Jacques has a new paper out on the subject of the Proto-Sino-Tibetan person-marking debate, a subject which I’ve been meaning to run a series on for some time. To summarize, the question of the PST person-marking debate is whether such a thing existed. Sino-Tibetan historical linguistics is not as complete as Indo-European historical linguistics, which over the course of about a hundred and fifty years progressed toward its current metastable state, where everybody is pretty sure that the current consensus reconstruction of Proto-Indo-European must be wrong on a fundamental level but nobody has any idea how to fix it. In fact, the internal cladistic structure of Sino-Tibetan is still controversial.

As it would be. The most potentially conservative morphology is found in the Rgyalrongic (also known as rGyalrong, Gyalrong, Gyarong, Jiarong, etc., a state of affairs that’s still better than that of its constituents: could you guess that Lavrung and Khroskyabs are the same language?) and Kiranti branches, which LaPolla, the main proponent of the view that PST had no verbal person-marking, says are related at a more recent level (the ‘Rung group’), and which Jacques, the main proponent of the view that it did, says are not. If PST had verbal person-marking, Rungic probably isn’t real; if Rungic is real, PST probably didn’t have verbal person-marking; if PST didn’t have verbal person-marking, Rungic probably is real; and if Rungic is real, PST probably didn’t have verbal person-marking.

Most of the evidence bearing on the debate comes from Rgyalrongic and Kiranti, but some comes from Tangut, Old Tibetan, and Old Chinese.

In Tangut, the person-marking suffixes are identical to the pronouns: the first-person singular pronoun and the first-person singular marker are both nga2, the second-person singular pronoun and the second-person singular marker are both nja2, and the second-person plural pronoun and the first- or second-person plural marker are both nji2Nga2 and nga2 are even written with the same character, although the other person markers have characters distinct from the pronouns.

However, the second-person plural pronoun nji2 can be used as an honorific singular, as in Old Tibetan and Early Modern English—in which case the verb takes the second-person singular marker nja2. If, that is, the verb is intransitive or has a third-person object: Tangut, you see, does not have accusative verbal person-marking.

If a Tangut verb is intransitive, it agrees with the agent of the verb. If it’s transitive and has a second-person patient (in which case it can’t have a second-person agent), it takes -nja2; if the patient is first-person, it takes -nga2, and if the patient is third-person, it takes –nga2 if the agent is first-person singular, -nja2 if the patient is second-person singular, and -nji2 if the patient is first- or second-person plural. (Third-person arguments are never marked.)

Even more confusingly (and more interestingly for our purposes), there’s a closed class of verbs in Tangut that have two stems, which of course are called Stem A and Stem B. Stem B is only used for first- or second-person singular reflexive verbs or verbs with a first- or second-person singular agent and a third-person patient. For example:

sjij1 mjo2 gja1 .jij1 rjyr2.wjij1 ljij2 mjy1djij2 lhjwo1djij2 mjy1tchjy1lji2nga2
“Today I see the army leave, but I will not see it return.”

.jwar1 gja1 ghu1 .jij1 ljiij2 lja1 zji.j1 kjij1ljij2nga2
“When the Yue army comes to destroy Wu, it will see me.”

This doesn’t look like a difference—in fact, in the conventional transcription of Tangut (which I’ve ASCIIfied slightly), with its bizarre and unnecessary proliferation of js, it looks like a typo—but the conventional transcription of Tangut is thankfully always accompanied by character numbers, which in this case are 4803 for the stem A lji2 and 0046 for the stem B ljij2; and Marc Miyake once reconstructed them as the more sensible 2lɨi and 2le. (He writes the tone before the syllable.)

This state of affairs must be conservative to some degree, but nobody knows where it came from.

Phonemic analysis is hard, let’s go shopping

Alexis Michaud’s phonology of Laze makes the unusual decision of providing a syllable chart instead of a phoneme inventory.

As it turns out, this is a sensible decision. A phonemic analysis of Laze, an account of its contrastive consonants and vowels (or, in the more applicable Sinitic tradition, onsets and rhymes) is difficult to produce, for the following reasons:

1) There are eleven possible syllables with nasal rhymes, of which eight begin with /h/ and three have a null initial: [hwã hɑ̃ hæ̃ hṽ̩ hũ hwɤ̃ hĩ hĩe æ̃ ṽ̩ ɔ̃]. The first eight can be analyzed as beginning with a nasal glottal fricative /h̃/, but the last three cannot. The rhymes and -ɔ can’t take a null initial, so [æ̃ ɔ̃] could be analyzed as /æ ɔ/—but [ṽ̩] contrasts with [v̩]. Michaud proposes to analyze these as h- or the null initial combined with nasal rhymes -wã -ɑ̃ -æ̃ -ṽ̩ -ũ -wɤ̃ -ĩ -ĩe -ɔ̃, under which analysis Laze would have six rhymes that can only occur after /h/, two that can only occur after /h ∅/ (∅ is the null initial, not to be confused with ø, the letter O WITH STROKE), and one that can only occur after /ø/.

It’s unclear from Michaud’s paper, but these last three syllables may be hapax phonoumena: he gives one example for [æ̃], two for [ṽ̩], and four for [ɔ̃], of which two are “likely Pumi/Prinmi loanwords”.

2) The initial [f] can only appear in native vocabulary before the rhymes [i] and [v̩]. [f] doesn’t contrast with [h] before [v̩], but it does before [i].

3) The alveolopalatal affricates [tɕʰ tɕ dʑ] appear to be allophones of the velars before front vowels or rhymes beginning with -j-; but in one word, [wɤ11mie11tɕɔ11ɭɔ55] ‘cicada’, an alveolopalatal affricate appears before a back vowel, probably due to Laze’s process of sporadic vowel copying. (Cf. [ʂieliemie] ‘seventh month’, from [ʂɯ] ‘seven’ + [ɬiemie] ‘month’.) There is no other -jɔ rhyme anywhere in Laze. Either /tɕ/ or /jɔ/ must be granted phonemicity to deal with this word.

4) The rhyme [e] can only appear after /s z tsʰ ts dz/, and “could actually be said to be in complementary distribution with any of the following rhymes: /ɤ/, /wɤ/, /ɹ̩/, /jæ/, /jɤ/, /jɑ/, /ɔ/, /ie/, /wæ/, /wɑ/, and even /æ/.” This is for diachronic reasons: *-a raised to -e after coronal initials and nowhere else, and there is no other source of -e. Michaud proposes that, since -wɤ is sometimes realized as close to [we], [e] could be an allophone of /ɤ/.

5) The rhyme [ɥe] appears in only two syllables, [ɮɥe] and [tɕɥe], each of which appears in only one word. [ɮɥe11] ‘wood ashes’ is likely a loan.  [tɕɥe55] “may likewise call for a special explanation, but so far we have not been able to clarify this point.” Michaud proposes to treat [ɥe] as an allophone of [wɤ], but [wɤ] can only appear after the null initial, /h/, velars, and retroflexes.

6) The uvular plosive initials [qʰ q] are in complementary distribution with the velars [kʰ k], only occurring before [ɑ wa æ ɔ], where [kʰ k] do not. However, there are syllables [kʰu ku], and [u] and [ɔ] only contrast in the syllable pairs [hu xɔ], [kʰu qʰɔ], [ku qɔ], [ɮu ɭɔ], and [bu bɔ]—and [bɔ] can only appear as a product of vowel harmony. (Possibly only in [bɔ33ɭɔ55] ‘fly’.) In fact, these and [ʁɔ] (and [ɔ̃] and [tɕɔ]) are the only syllables that [ɔ] can appear in.

The uvulars could be analyzed as allophones of the velars before low vowels, but there’s also a uvular fricative [ʁ], which appears in almost-complementary distribution with [g]: [ʁ] can only appear before [ɑ æ wæ wɑ ɔ] (no other velars or uvulars can appear before [wæ]), but, due to vowel harmony, [g] can appear before [ɑ] and [wæ].

7) Retroflex and alveolar plosives only contrast before [v̩].

8) The retroflex lateral [ɭ] only appears before the marginal phoneme [ɔ]. In the closely related Naxi language, the syllable [ɭɔ] is analyzed as /lo/, since retroflexes and alveolars only contrast before /o/; but Laze has no /o/, and [ɔ] is marginal. Michaud proposes to recognize /ɔ/ as a phoneme and analyze this syllable as /lɔ/, creating a final that can only appear after x-, uvulars, b- (due to vowel harmony), and l-.

9) The distribution of most segments is severely restricted. v-, for example, can only appear before three (possibly four) rhymes, and -wɑ can only appear after seven initials.

Hapax phonotactica

In addition to phonemes that only occur in one or a few words, there are can be unusual exceptions to a language’s usual phonotactics: consonants that only occur in a certain position in one word, clusters that only occur in one word, and so on.

Latin /p/ only occurs finally in volup, a shortened form of volupis.

German -nf only occurs in Hanf ‘hemp’ (a pre-PGmc. loan of Latin cannabis), Senf ‘mustard’ (< Latin sināpi), Genf ‘Geneva’ (< Latin Genava), Sernf (a river; < *Sarnivos), and fünf ‘five’ (< *pempe, a variant of PIE *penkʷe).

English oi only precedes a word-final velar in oikyoik (a loan from Finnic), and the onomatapoeia oinkboink, and boing. English can only appear in initial clusters in loans and recent coinages: VladimirvlogVlachvlei.

But these are less interesting: it’s well-known that loanwords and onomatapoeia can fill in phonotactic gaps in a language.

Sometimes these clusters can change in unusual ways: Old English fn- (in fnēosan ‘sneeze’) becomes sn-, possibly aided by the visual similarity of and long s.

Hapax phonoumena

(edit: changed title as per Cev’s recommendation; there ought to be a standard term for these)

Some languages have contrastive phonemes that appear in only one root: for example, Norwegian /ʉi̯/ only appears in the word /hʉi̯/ hui and the derived verb huie; Arabic /lˤ/ only appears in /alˤlˤah/ ‘Allah’; Spanish /ui̯/ only appears in muy; and in standard American English, unconditioned /eə/ only occurs in the interjection ‘yeah’. (Interjections, of course, often have expanded phoneme inventories relative to the rest of the language—see English oink and boink, which have /oi/ before a non-coronal cluster—but these other two examples are of ordinary roots.)

According to Wikipedia, Dahalo has five such phonemes: /ⁿd̠ʷ/, /ᶮdʒ/ in /kípuᶮdʒu/ ‘place where maize is seasoned’, /ᵑɡʷ/ in /háᵑɡʷaraᵑɡʷára/ ‘centipede’, /ɬʷ/, in /ɬʷaʜ-/ ‘to pinch’, and /j/, in /jáːjo/ ‘mother’.

According to Blust’s Austronesian Languages, unconditioned nasal vowels only appear in one root in Bintulu and Miri: Bintulu has the negation marker [ʔã] ã and Miri has the minimal pair haaw ‘rafter’ : hããw ‘2sg’, but no other unconditioned nasal vowels are known in either language.

Mako /ə/ only appears in the past-tense suffix -tə, but it’s fully contrastive: every vowel can appear word-finally after /t/.

Qiang /ɦ/ only occurs in the interjection /ɦa/ and a directional prefix. (There are many other hapax phonoumena in the Qiangic languages; I won’t list them all.)

Hoyahoya /s/ only occurs in /sa/ sha ‘bone’. (/s/ is written sh because the other fricatives, /ʁ/ and /h/, are written gh and h.)

The Amuzgo syllabic velarized prenasalized bilabial trill only occurs in the word [ʃa˥m̩ˠʙˠ˥] ša1ṃb1 ‘antlion’.

Mianchi /ɬ/ occurs only in /ɬə̀/ ‘moon’ ( < PTB *s-la), and only for some speakers; others have /l/.

In some dialects of Mandarin, the final -iai only occurs in 崖 yái ‘precipice’.

Sometimes these phonemes are supported by loanwords.

Japhug /y/ occurs in Chinese loanwords, but only in one native root, /qaɟy/ ‘fish’.

Yadu /æ/ occurs in Chinese loanwords and one native root, /tsæm/ ‘girl’.

In Longxi, /h/ occurs in two loanwords, an onomatopoeic form, and /hàN háN/ ‘corridor’; and /v/ occurs only in words derived from Proto-Qiang *u ‘you’: /vù/ ‘you’ and its compounds, /vú lià/ ‘we (incl.)’ and its dual, and /vé ì/ ‘yourself’ and its dual and plural; however, /v/ may also be analyzed as an allophone of /u/ in word-initial position (pure [u] never appears word-initially), making these forms /ù/, /ú lià/, and /ué ì/, with a u-deletion rule for the reflexive.

Sierra Miwok /š/ occurs only in English loanwords and the exclamation /ʔiš·o·/ ‘Scat!’.

Sometimes they occur in several roots. Chechen /r̥/ only occurs in /vuor̥/ ‘seven’ and /bar̥/ ‘eight’. The closely related language Bats has /vorɬ/ and /barɬ/.

Sometimes they are supported by the morphology. In Finnish, the diphthong ey merged into öy except in the verb leyhyä, but it can still occur in derived words, when e- is adjacent to -U in a front-harmonic context.

What were Old Chinese A and B syllables?

I don’t know. Neither does anyone else. There are innumerable proposals. All are contradictory, and some are the exact reverse of others: Pulleyblank (1962) and Zhengzhang (1987) both attributed the A/B contrast to vowel length, but Pulleyblank thought type-B vowels were long, and Zhengzhang thought type-A vowels were.

…But I analyzed the 4967 (at least, I hope that’s what the count is!) Old Chinese items from the recent Baxter-Sagart reconstruction, to find out how often each vowel occurred in each type of syllable.

Vowel A B B – A
*a 810 789 -21
*e 345 340 -5
*i 155 305 150
*o 319 344 25
*u 229 358 129
301 607 306
*A 0 65 65

I don’t know what *A is. I also don’t know what this distribution suggests. High vowels, especially *ə, seem to prefer type-B syllables, but low vowels don’t care one way or the other.

Amritas‘s old proposal that the A/B distinction could have come from low/high presyllabic vowels reminded me to check presyllables:

Presyllable type A B
None 1687 2073
Tightly-bound 239 351
Loosely-bound 230 380
Both 3 4

In the corpus, there are 2159 type-A syllables and 2808 type-B syllables; that is, 56.53% of syllables are type-B. There are 1200 words with one presyllable; if there’s no correlation, we’d expect about 678 of them to be of type B. In fact, there are 731. This is probably not significant. The same for words with no presyllable: we’d expect 2125, and there are 2073. Unless Proto-Sinitic had an Austronesian-like CVCVC word structure and lost most initial consonants—that is, unless it was Arrernte* or something—the presyllabic vowel hypothesis is probably out.

I wonder what the statistical distribution of pharyngealization (or similar) looks like in the Qiangic languages that have it. Incidentally, where did that come from? I think Guillaume Jacques once said in a paper that it’s unknown. Could it have been preserved from Proto-Sino-Tibetan?

* Arrernte underwent the historical change of unconditionally dropping all word-initial consonants.

Conscious manipulation of the phonetic substance

A number of the languages of Borneo also have preploded final nasals, which have arisen from word-final simple nasals in syllables that have a non-nasal onset. These are perhaps best known from the Land Dayak-Kendayan ([kənᵈájaᵗn]) Dayak area of southern Sarawak and adjacent parts of the Indonesian province of West Kalimantan, but they have a geographical distribution which is very similar to that of postploded medial nasals. The term ‘nasal preplosion’ was coined by Court (1967) in a brief description of the phonology of Mentu Land Dayak. In this language nearly all final nasals are preceded by a brief voiceless oral onset, as in əsɨᵖm‘sour’, burəᵗn ‘moon’, or turaᵏŋ ‘bone’. The exceptions fall into two classes: personal names, and bases in which the final syllable begins with a nasal consonant. While the first class of exceptions evidently is motivated by conscious manipulation of the phonetic substance for purposes of marking semantic fields, the second is due to purely phonetic factors. Phonetic information on vowel nasality is poorly reported for most AN languages, but relatively good data is available for the languages of Borneo, and in these it is clear that vowels are most strongly nasalised by a preceding primary nasal consonant in a process that can be described as ‘onset-driven’ nasal harmony (Blust 1997c). Coda-driven nasal harmony is generally absent, but some minor leakage of nasality into vowels that precede a final nasal must take place in most languages, or final nasals would be universally preploded. The preplosion of final nasals can therefore be seen as a strategy for blocking nasal spreading in the ‘wrong’ direction. When a final syllable begins with a nasal consonant this is impossible, as in Mentu Land Dayak inəm ‘six’. The oral element in a preploded nasal is voiceless in some languages, such as Mentu Land Dayak, voiced in others (as Bau/Senggi Land Dayak), and mixed in Bonggi, spoken on Banggi Island north of Sabah, where Boutin (1993:111) reports -ᵇm, -ᵈn, but -ᵏŋ.

Robert Blust, The Austronesian Languages